On Sample Based Explanation Methods for NLP: Faithfulness, Efficiency and Semantic Evaluation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 489–502 August 1–6, 2021. ©2021 Association for Computational Linguistics 489 Generating Landmark Navigation Instructions from Maps as a Graph-to-Text Problem Raphael Schumann Computational Linguistics Heidelberg University, Germany Stefan Riezler Computational Linguistics & IWR Heidelberg University, Germany {rschuman|riezler}@cl.uni-heidelberg.de Abstract Car-focused navigation services are based on turns and distances of named streets, whereas navigation instructions naturally used by hu- mans are centered around physical objects called landmarks. We present a neural model that takes OpenStreetMap representations as input and learns to generate navigation in- structions that contain visible and salient land- marks from human natural language instruc- tions. Routes on the map are encoded in a location- and rotation-invariant graph repre- sentation that is decoded into natural language instructions. Our work is based on a novel dataset of 7,672 crowd-sourced instances that have been verified by human navigation in Street View. Our evaluation shows that the nav- igation instructions generated by our system have similar properties as human-generated in- structions, and lead to successful human navi- gation in Street View. 1 Introduction Current navigation services provided by the au- tomotive industry or by Google Maps generate route instructions based on turns and distances of named streets. In contrast, humans naturally use an efficient mode of navigation based on visible and salient physical objects called landmarks. As shown by Tom and Denis (2004), route instructions based on landmarks are easier processed and mem- orized by humans. May et al. (2003) recommend that in pedestrian navigation systems, ”landmarks should be used as the primary means of providing directions”. Another navigation scenario where landmarks are useful is if GPS tracking is poor or not available, and if information is inexact regard- ing distances (e.g., in human estimates) or street names (e.g., for users riding a bicycle). We present a neural model that takes a real-world map repre- sentation from OpenStreetMap1 as input and gen- erates navigation instructions that contain salient landmarks, learned directly from human natural language instructions. In our framework, routes on the map are learned by discretizing the street layout, connecting street segments with adjacent points of interest, thus en- coding visibility of landmarks, and encoding the route and surrounding landmarks in a location- and rotation-invariant graph. Based on crowd-sourced natural language instructions for such map rep- resentations, a graph-to-text mapping is learned that decodes graph representations into natural lan- guage route instructions that contain salient land- marks. Our work is accompanied by a dataset of 7,672 instances of routes in OpenStreetMap and corresponding crowd-sourced natural language in- structions. The navigation instructions were gen- erated by workers on the basis of maps including all points of interest, but no street names. They were verified by different workers who followed the navigation instructions on Google Street View2. Experimental results on randomly sampled test routes show that our graph-to-text model produces landmarks with the same frequency found in hu- man reference instructions. …

Content truncated to 500 words